## _wp06289

## Source details

**Canonical URL:** [_wp06289](https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06289.pdf)

## Other formats

- [Markdown version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06289.pdf.md)
- [Structured JSON version](/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06289.pdf.json)

---

### I. Introduction: purpose and scope
- Objective: assess the joint hypothesis that (a) transparency improves the accuracy of private sector forecasts and (b) inflation targeting (IT) enhances transparency.
- Transparency definition (Geraats, 2002): "the removal of information asymmetries."
- Five elements of transparency considered: political, economic, procedural, policy and operational.
- Claimed IT transparency channels:
  - Political transparency via specification of target variable and precise numerical target.
  - Procedural, policy, and operational transparency via enhanced policy analysis, openness about internal deliberations, and greater explanation of decisions.
  - Economic transparency via explicit/public discussion of forecasts, attendant risks, and underlying assumptions.
- Preview of empirical conclusions:
  - Strong evidence for MS prediction (i): IT enhances transparency particularly for those with poor private information.
  - No evidence for MS prediction (ii): no indication that IT harms the best private forecasters.

### Theoretical background and simple signal-extraction model
- Traditional model insights: greater transparency can reduce inflation bias but may limit ability to offset output surprises; costs of transparency require other distortions.
- Morris and Shin (2002) (MS) key insight: public information can crowd out high-quality private information; improved public information may help poor forecasters and can, in principle, harm the best forecasters.
- Simple signal-extraction model (Section II) — preserved equations and definitions:
  - Agent objective: u_i(f_i; π) = −(f_i − π)^2 (equation (2))
  - Public signal: C = π + η (equation (3))
  - Private signal: i = π + ε (equation (3) continued)
  - Precision of public signal: ρ = 1/σ_η^2 (equation (4))
  - Precision of private signal: φ = 1/σ_ε^2 (equation (4) continued)
  - Optimal forecast: f_i* = (ρ C + φ i) / (ρ + φ) (equation (5))
  - Expected mean square forecast error: E[(f_i − π)^2] = 1/(ρ + φ) (equation (6))
- Identification and IT effect:
  - Private-signal precision constant: φ_it = φ_i
  - Public-signal precision depends on country and IT: ρ_jt = ρ_j(IT) (equation (7))
  - IT improves transparency iff ρ_j(1) > ρ_j(0)
  - In non-IT case: eV_ij|IT=0 = eV_ij^0 = 1/(ρ_j(0) + φ_i) (equation (8))
  - Comparative statics: ∂ eV_ij / ∂ ρ_j = −1/(ρ_j(0) + φ)^2 = −(eV_ij^0)^2 < 0 (equation (9)); ∂^2 eV_ij / ∂ ρ_j ∂ eV_ij^0 = −2 eV_ij^0 < 0 (equation (10))
  - Linearized IT effect: Δ eV_ij ≡ eV_ij^1 − eV_ij^0 ≈ −b_0 − b_1 eV_ij^0 with −b_1 < 0 (equation (11))
  - MS mapping to empirical coefficients:
    - Prediction (i): −b_1 <0 — worst forecasters (higher V_0) experience largest improvement.
    - Prediction (ii): b_0 >0 — best forecasters (V_0 ≈0) could see errors increase (requires additional parameter restrictions per MS).

### Empirical strategy and estimation
- Empirical equation (difference in per-forecaster error before/after IT):
  - ΔV_ij ≡ V_ij^1 − V_ij^0 = −b_0 − b_1 V_ij^0 + Δ e_ij | jIT=1 − Δ e_ij | jIT=0 (equation (12))
- Estimation regression (OLS specification):
  - ΔV_ij = b_0 + b_0^T D_ij^T − b_1 + b_1^T D_ij^T V_ij^0 + u_ij (equation (13))
  - D_T is dummy for treatment (IT adoption)
- Mean reversion and endogeneity concerns:
  - Cov(V_ij^0, Δ e_ij) < 0 ⇒ cov(V_ij^0, u_ij) < 0 ⇒ negative bias on coefficient of V_ij^0
  - Include −b_1 V_ij^0 to control for mean reversion (following Ball and Sheridan (2004))
- Instrumental variables (IV) to address remaining endogeneity:
  - Instruments for V_ij^0:
    - V_ij^0,g: forecaster i's mean absolute forecast error for GDP growth (next year)
    - f_ij: expected inflation rate for the one-year period prior to IT adoption date
  - Instruments for D_T V_ij^0:
    - D_T V_ij^0,g and D_T f_i
- Estimation methods: OLS and IV/2SLS (two-stage least squares)

### Data, windows, and measurement
- Data source: Consensus Forecasts dataset (private-sector monthly or bimonthly forecasts)
- Focus: “next year” inflation forecasts (following Johnson, 2002)
- Country and episode coverage: eleven IT-adoption episodes — four industrial countries (Australia, Canada, Norway, United Kingdom) and seven emerging markets (Brazil, Chile, Colombia, Korea, Mexico, Peru, Thailand)
- Adoption dating: month and year from Roger and Stone (2005)
- Panel construction: 24-month window centered on IT adoption (12 months before and 12 months after)
- Outcome measure: average (per-forecaster) absolute forecast error (absolute difference between next-year inflation forecast and actual next-year annual inflation from IMF International Financial Statistics) used as V_ij^0 and V_ij^1
- Treatment sample size: 166 forecasters with before-and-after data in the 11 IT-adopting countries
- Measurement caveats: V_ij contaminated by idiosyncratic time-varying shocks e_ijt and measurement error; addressed by regression controls, IVs, and matching

### Matching treatment and control groups (Appendix II summary)
- Control pool: forecasters in countries that did not adopt IT during the same 24-month period (or following 12-month period) and had not adopted IT previously
- Propensity score probit uses eight variables:
  - V_ij^0,g; V_ij^0; ΔV_ij−1,g; ΔV_ij−1; f_ij^0,g; f_ij^0; Δf_ij−1,g; Δf_ij−1
- Matching methods (three algorithms):
  - Algorithm 1 (preferred): nearest available control with replacement (random ordering; allows reuse of controls)
  - Algorithm 2: one-to-one matching without replacement
  - Algorithm 3: two-step "best country" matching (first draw best matches; then resample controls from country(ies) providing highest number of matches)
- Common support enforcement and balance checking performed
- ps estimation: total observations 2,141; probit χ2(8)=87.6 [p-value=0.000]; pseudo-R2 = .07
- Table 1 frequency-weighted descriptive statistics (selected):
  - Observations (Unweighted): 166, 109, 86 (Panel counts)
  - V_ij0 (mean (sd)):
    - Treatment: 2:14 (1:58)
    - Control 1: 2:93 (4:09)
    - Control 2: 3:14 (4:35)
    - Control 3: 3:78 (5:06)
  - ΔV_ij (mean (sd)) — formatted as in source:
    - Treatment: V_ij :934 (1:45)
    - Control 1::237 (3:15)
    - Control 2::0697 (3:54)
    - Control 3: :0465 (4:04)

### Regression specifications and reporting conventions (Table 3)
- Three panels correspond to matching algorithms:
  - Panel A: Nearest Neighbor Matching (with replacement)
  - Panel B: One-for-One Nearest Neighbor Matching (without replacement)
  - Panel C: "Best country" Nearest Neighbor Matching (with replacement)
- Three model variants:
  - Model I ("Levels"): suppresses interaction effect; measures unconditional IT effect on forecast errors
  - Model II ("Interaction (OLS)"): full equation estimated by OLS
  - Model III ("Interaction (2SLS)"): IV/2SLS using instruments V_ij^0,g and f_ij
- Residual treatments reported per specification: iid, robust, clustered
- Advanced-country dummy included; advanced countries listed as Australia, Canada, France, Germany, Italy, Japan, the Netherlands, Norway, Spain, Switzerland, the United Kingdom, and the United States
- Selected coefficient estimates (preserved formatting and numeric values as reported):
  - Panel A highlights:
    - IT (Levels / Interaction columns):  1:16 yyy (:272),  1:16 yyy (:305),  1:16  (:666)
    - V_ij0: 0 :153 (:0480),  :153 (:0874),  :153 (:124)
    - IT × V_ij0:  :752 yyy (:119),  :752 yyy (:0891),  :752 yyy (:154)
    - F_stat entries: 9:49 yyy, 23:0 yyy, 2:47 
    - R2 entries: :0545, :0545, :0545; :216, :216, :216; :193, :193, :193
  - Panel B highlights:
    - IT:  :961 yyy (:300),  :961  (:336),  :961 (:729)
    - V_ij0:  :286 yyy (:0491),  :286 yyy (:0882),  :286  (:112)
    - IT × V_ij0:  :637 yyy (:128),  :637 yyy (:0891),  :637 yyy (:152)
    - F_stat entries: 6:24 , 23:0 yyy, 2:37
    - R2 entries: :0366, :0366, :0366; :225, :225, :225; :210, :210, :210
  - Panel C highlights:
    - IT:  :827  (:335),  :827  (:377),  :827 (1:26)
    - V_ij0:  :365 yyy (:0477),  :365 yyy (:0631),  :365  (:107)
    - IT × V_ij0:  :598 yyy (:139),  :598 yyy (:0705),  :598  (:190)
    - F_stat entries: 4:47 , 23:6 yyy, 2:22
    - R2 entries: :0265, :0265, :0265; :246, :246, :246; :223, :223, :223
- Significance notation: "SigniÖcance levels in this and subsequent tables are denoted as: yyy" (table truncates after that label in provided content)

### Main empirical findings (selected quantitative results)
- Strong evidence for the conditional effect predicted by the model (columns II and III), but little evidence for an unconditional effect (column I).
- OLS estimate for b1T (column II) ranges between 0.6 (panel C) and 0.75 (panel A).
- Using clustered standard errors, OLS estimates are highly statistically significant: at least at the 1% level (panel C) and generally at the 0.1% level (panels A and B).
- IV/2SLS estimates (column III): point estimate for b1T between 0.8 and 0.94; statistical significance level similar to OLS.
- IV strategy mitigates mean reversion bias: IV estimate of b1 is not significantly different from zero whereas OLS estimate is highly significant and larger in magnitude.
- Column I (unconditional effect): not generally statistically significant and carries a negative sign in all specifications.
- Placebo tests (Table 4): shifting IT adoption window 12 months earlier or 12 months later yields no statistically significant conditional or unconditional effect from the placebo IT adoption variable (first adoption episode, Canada, drops out due to data constraints).
- Robustness checks (Table 5):
  - Trimming controls with the highest 5 percent of V_ij0, or dropping controls with V_ij0 above treatment-group maximum (7.4), has little effect on the measured conditional effect of IT.
  - IV estimates on restricted samples often strengthen point estimates and significance for b1T.
  - Dropping Venezuela reduces point estimate on b1T but coefficient remains significantly different from zero except in panel C (columns VIII and IX).
  - One case reports an IV p-value of .122 (borderline significance).
- Standard errors and clustering:
  - Three sets of standard errors reported: iid, robust, clustered (clustered preferred).
  - Number of clusters for clustered residuals:
    - Panel A: 58 clusters.
    - Panel B: 63 clusters.
    - Panel C: 23 clusters.
- IV diagnostics (clustered residuals) reported to pass Anderson canonical correlations LR statistic, Hansen's J statistic, and Stock and Yogo weak-instrument tests (noting homoskedasticity requirement caveat).

### Quantitative sample reporting
- Observations (weighted) reported in Table 4: 272 272 272 352 352 352.
- Table 5 Obs. (Clusters) reporting examples:
  - Panel A: Obs. (Clusters) 307 (54), 304 (54), 305 (53)
  - Panel B: Obs. (Clusters) 316 (62), 308 (61), 302 (57)
  - Panel C: Obs. (Clusters) 306 (21), 298 (21), 294 (20)

### Interpretation, policy implications, and caveats
- Empirical conclusion: adoption of IT leads to greater convergence (stronger mean reversion) in inflation forecast errors, consistent with the signal-extraction model and MS prediction (i) that better public information helps forecasters with poorer private signals most.
- No evidence that IT raises forecast errors for the best forecasters (no support for MS prediction (ii) in the data).
- Policy implication: findings constitute compelling evidence that IT enhances transparency in inflation forecasting — relevant to policymakers concerned with expectations and the monetary transmission mechanism.
- Caveat and further research: additional tests could examine IT's transparency effect for other variables; preliminary results suggest little or no effect for output growth forecasting.

*Source: _wp06289 (excerpted content).*

### References..............................................................................................................

### _wp06289 - References

### I. Introduction: purpose and scope
- Objective: assess the joint hypothesis that (a) transparency improves the accuracy of private sector forecasts and (b) inflation targeting (IT) enhances transparency.
- Definition of transparency (following Geraats, 2002): "the removal of information asymmetries."
- Five areas of monetary policymaking where transparency could affect outcomes: the central bank's objectives and institutional relationship with government; publication of data and forecasts; internal decisionmaking; communication and explanation of policy changes; and details of policy implementation — yielding five elements of transparency (political, economic, procedural, policy and operational).
- Claimed IT transparency channels:
  - Political transparency via specification of target variable and precise numerical target.
  - Procedural, policy, and operational transparency via enhanced policy analysis, openness about internal deliberations, and greater explanation of decisions.
  - Economic transparency via explicit/public discussion of forecasts, attendant risks, and underlying assumptions.

### Theoretical background on transparency and forecasting
- Traditional models (Barro-Gordon (1983), Kydland-Prescott (1977), Cukierman and Meltzer (1986)) suggest:
  - Greater transparency can reduce inflation bias but may limit the central bank's ability to offset output surprises.
  - Costs of transparency require existence of distortions in addition to information asymmetry.
- Alternate views emphasize long-term goals and credibility (King, 1997).
- Morris and Shin (2002; MS) dissent: public information can act as a focal point for "beliefs about beliefs" and may crowd out high-quality private information, potentially making private forecasts more variable.
  - MS show (Appendix I summary):
    - Enhanced transparency is most likely to benefit forecasters with less accurate private signals.
    - The best forecasters can, in principle, be harmed by more transparent public information.

### Hypotheses and simple model representation
- Empirical test framed via the effect of enhanced transparency on forecast errors as the sum of:
  - a constant effect for all forecasters, and
  - an interaction effect that falls with forecast accuracy.
- Model expression (as in source):
  - V=b 0  b 1 V 0 :(1)
- MS predictions mapped to coefficients:
  - Prediction (i): −b 1 <0 — worst forecasters (higher initial forecast errors V 0) experience the greatest improvement.
  - Prediction (ii): b 0 >0 — best forecasters (those with V 0 ≈0) see forecast errors increase when public information improves.
- Section II outlines a simple signal-extraction model (a special case of MS with the beauty-contest weight set to zero) where equation (1) is explicitly derived.
  - In that simple model −b 1 <0 is consistent.
  - b 0 >0 requires additional parameter restrictions and is particular to MS (discussed in Appendix I).

### Empirical context and related literature
- Expected benefits of IT on forecasting:
  - Enhanced transparency should make private-sector forecasts of variables influenced by the central bank more accurate.
- Existing empirical evidence (survey):
  - Chortareas, Stasavage, and Sterne (2002): greater transparency (more prominence for central bank forecasts) associated with lower inflation in a diverse cross-section — may reflect credibility benefits.
  - Romer and Romer (2000): unpublished Fed forecasts are superior to professional private forecasts; if published they might dominate private forecasts (but publication could change accuracy).
  - Eijffinger and Geraats (2006): Fed's political transparency weaker than comparable institutions, policy transparency compares favorably with IT institutions.
  - Swanson (2004): financial market forecasts of policy interest rates in the United States have improved with Fed transparency changes.
  - Gurkaynak, Levin, and Swanson (2005): in UK and Sweden IT makes long-term inflation expectations less responsive to economic "news", suggesting better anchoring of expectations (could reflect credibility, not necessarily transparency).
  - Corbo, Landerretche and Schmidt-Hebbel (2001): drop in forecast errors among IT adopters, though decline often predates IT adoption.
  - Johnson (2002): panel of 11 industrial countries — some evidence for credibility benefits of IT (expected inflation falls more in targeters) but no evidence for transparency benefits (no reduction in forecast variability or absolute error).

### Preview of results reported in the paper (as stated in text)
- Strong evidence for MS prediction (i): IT enhances transparency particularly for those with poor private information (worst forecasters improve).
- No evidence for MS prediction (ii): no indication that IT harms the best private forecasters.
- Interpretation: IT enhances transparency and is especially helpful for forecasters with poor private signals, while it is unlikely to harm precise private forecasters.

### Methodological and measurement notes
- Cautions on transparency measurement:
  - Subjective transparency indices (e.g., Fry and others, 2000; Eijffinger and Geraats, 2006; Roger and Stone, 2005) risk tautological bias if indices incorporate features closely linked to IT.
  - Use of survey-derived measures can introduce additional biases.
- Relevance of Fed experience:
  - Debate over publication of voting records/transcripts and delayed release effects on incentives (Meade, 2006).
  - Media narratives around appointments underscore perceived differences in depersonalization of policy under IT versus non-IT regimes.

*Source: _wp06289 - References (excerpt).*

### Section II outlines a simple signal extraction model with public and private information to

### _wp06289 - Section II outlines a simple signal extraction model with public and private information to

### Theoretical framework: simple signal-extraction model
- Objective: Agents (forecasters) minimize squared forecast error:
  - u_i(f_i; π) = −(f_i − π)^2 (equation (2))
- Information structure:
  - Public signal: C = π + η (equation (3))
  - Private signal: i = π + ε (equation (3) continued)
- Signal precision definitions:
  - Precision of public signal: ρ = 1/σ_η^2 (equation (4))
  - Precision of private signal: φ = 1/σ_ε^2 (equation (4) continued)
- Optimal weighting of signals by forecaster i:
  - f_i* = (ρ C + φ i) / (ρ + φ) (equation (5))
- Expected mean square forecast error (forecast inaccuracy):
  - E[(f_i − π)^2] = 1/(ρ + φ) (equation (6))
- Identifying assumptions:
  - Private-signal precision is constant over time for each forecaster i: φ_it = φ_i
  - Public-signal precision depends only on country-specific factor and IT adoption (IT ∈ {0,1}):
    - ρ_jt = ρ_j(IT) (equation (7))
  - IT improves transparency iff ρ_j(1) > ρ_j(0)
- Implication for forecast error in non-IT (IT=0) case:
  - eV_ij|IT=0 = eV_ij^0 = 1/(ρ_j(0) + φ_i) (equation (8))
- Comparative statics:
  - ∂ eV_ij / ∂ ρ_j = −1/(ρ_j(0) + φ)^2 = −(eV_ij^0)^2 < 0 (equation (9))
  - ∂^2 eV_ij / ∂ ρ_j ∂ eV_ij^0 = −2 eV_ij^0 < 0 (equation (10))
- Linearized effect of IT on forecast errors (approximation around eV_ij^0):
  - Δ eV_ij ≡ eV_ij^1 − eV_ij^0 ≈ −b_0 − b_1 eV_ij^0 with −b_1 < 0 (equation (11))
- Empirical strategy implied: difference-in-differences comparing forecasters in IT adopters (treatment) and non-adopters (control), exploiting interaction with baseline forecast inaccuracy.

### Empirical strategy and estimation approach
- Empirical counterpart acknowledges contamination by idiosyncratic time-varying shocks e_ijt and measurement error.
- Empirical equation (difference in per-forecaster error before/after IT):
  - ΔV_ij ≡ V_ij^1 − V_ij^0 = −b_0 − b_1 V_ij^0 + Δ e_ij | jIT=1 − Δ e_ij | jIT=0 (equation (12))
- Estimation regression (OLS specification):
  - ΔV_ij = b_0 + b_0^T D_ij^T − b_1 + b_1^T D_ij^T V_ij^0 + u_ij (equation (13))
  - D_T is dummy for treatment (IT adoption)
- Mean reversion and endogeneity:
  - Cov(V_ij^0, Δ e_ij) < 0 ⇒ cov(V_ij^0, u_ij) < 0 ⇒ negative bias on coefficient of V_ij^0
  - Include −b_1 V_ij^0 term to control for mean reversion (strategy follows Ball and Sheridan (2004))
- Instrumental variables (IV) approach proposed to address remaining endogeneity:
  - Instruments for V_ij^0:
    - V_ij^0,g: forecaster i's mean absolute forecast error for GDP growth (next year) analogous to V_ij^0 for inflation (instrument relevance: common forecaster ability; exogeneity: orthogonality of idiosyncratic shocks across variables) (footnote discussion)
    - f_ij: expected inflation rate for the one-year period prior to IT adoption date (uses expected rather than actual inflation to avoid correlation with unexpected inflation shocks)
  - Instruments for D_T V_ij^0:
    - D_T V_ij^0,g and D_T f_i
- Estimation methods used/tested:
  - Ordinary least squares (OLS)
  - IV estimation using the two instruments above

### Data, windowing, and measurement
- Data source: Consensus Forecasts dataset (private-sector monthly or bimonthly forecasts)
- Focus: “next year” inflation forecasts (following Johnson, 2002)
- Country and episode coverage:
  - Eleven IT-adoption episodes:
    - Four industrial countries: Australia, Canada, Norway, United Kingdom
    - Seven emerging markets: Brazil, Chile, Colombia, Korea, Mexico, Peru, Thailand
  - Some countries that adopted IT were not in the sample at adoption date
- Adoption dating: month and year from Roger and Stone (2005) (used as consensus dating)
- Forecast horizon and panel construction:
  - For each country a panel of forecasters with changing composition over time; identification focuses on individual forecasters
  - 24-month window centered on IT adoption (12 months before and 12 months after) is used to exploit monthly data and sharpen identification
  - Average (per-forecaster) absolute forecast error (absolute difference between next-year inflation forecast and actual next-year annual inflation from IMF International Financial Statistics) used as proxies for V_ij^0 and V_ij^1
- Sample used for treatment group:
  - 166 forecasters with before-and-after data in the 11 IT-adopting countries form the treatment group
- Measurement concerns:
  - V_ij will be contaminated by idiosyncratic time-varying shocks and measurement error; model incorporates error term e_ijt and addresses mean reversion via regression controls and IVs

### Matching treatment and control groups
- Purpose: eliminate selection bias in difference-in-differences
- Control pool: forecasters in countries that did not adopt IT during the same 24-month period (or the following 12-month period) and which had not adopted IT previously
- Propensity score specification (eight variables used):
  - Mean absolute forecast errors for output and inflation in the 12-month period up to IT adoption: V_ij^0,g and V_ij^0
  - Changes in these variables from period −1 to period 0: ΔV_ij−1,g and ΔV_ij−1
  - Forecast-level variables for expected inflation and growth: f_ij^0,g and f_ij^0
  - Changes in these forecast-level variables from period −1 to period 0: Δf_ij−1,g and Δf_ij−1
- Matching methods (three variants for robustness):
  - Preferred method: nearest available control with replacement (best available control for each treated observation; allows a control to be matched multiple times)
  - One-to-one matching without replacement (randomized ordering important)
  - Two-step country-resampling method:
    - First draw best matches (with replacement) from full control set
    - Identify country providing highest number of matches (ties: all selected)
    - Resample controls from that limited set according to propensity score (matching with replacement)
- Balance check:
  - Figure 1 (described) shows distribution of propensity score within treatment and three control groups; distributions are similar (observations frequency weighted for control groups 1 and 3)

### Key empirical identification challenges and solutions
- Mean reversion in forecast accuracy biases OLS estimates (negative correlation between baseline error V_ij^0 and change ΔV_ij)
- Control for mean reversion by including V_ij^0 term in regression (−b_1 V_ij^0)
- Use IVs (V_ij^0,g and f_ij) to isolate the fundamental component of forecaster ability from transient shocks
- Matching on observables via propensity score matching to construct comparable control group and mitigate selection into IT adoption confounding

*Source: _wp06289 - Section II outlines a simple signal extraction model with public and private information to*

### Appendix II provides a full account of the matching algorithms used in the paper. Table 1

### _wp06289 - Appendix II provides a full account of the matching algorithms used in the paper. Table 1

### Appendix scope and datasets
- Appendix II: "provides a full account of the matching algorithms used in the paper."
- Table 1: "provides summary statistics (means with standard deviations in parentheses) for the treatment and three control groups."
- Table 2: "The three samples of treatment and control groups are detailed in Table 2."
- Observations (Table 1):
  - Unweighted: 166, 109, 86
  - Weighted: 166, 166, 166, 166

### Table 1 — Descriptive statistics (frequency weighted)
- Group columns: Treatment, Control 1, Control 2, Control 3
- V_ij0 (mean (sd)):
  - Treatment: 2:14 (1:58)
  - Control 1: 2:93 (4:09)
  - Control 2: 3:14 (4:35)
  - Control 3: 3:78 (5:06)
- ΔV_ij (mean (sd)) — presented with signs and formatting as in source:
  - Treatment: V_ij :934 (1:45)
  - Control 1::237 (3:15)
  - Control 2::0697 (3:54)
  - Control 3: :0465 (4:04)

### Table 2 — Sample country coverage (Unweighted Observations)
- Panel A: Control Group 1 — totals and treatment/control counts by targeter country
  - Total (per targeter): Australia 25, Brazil 25, Canada 31, Chile 19, Colombia 16, Korea 24, Mexico 30, Norway 18, Peru 20, Thailand 19, United Kingdom 48, Total 275
  - Treatment (per targeter): 14, 14, 18, 13, 10, 15, 17, 10, 11, 11, 33, Total 166
  - Controls (per targeter): 11, 11, 13, 6, 6, 9, 13, 8, 9, 8, 15, Total 109
  - Additional country entries (counts as in table): Argentina 1, France 2, Germany 6, Hong Kong SAR 1, India 2, Indonesia 1, Italy 3, Japan 1, Malaysia 2, Netherlands 1, Singapore 1, Spain 1, Switzerland 1, United States 1, Venezuela 3 (and other per-cell counts as shown in the table)
- Panel B: Control Group 2 — totals and treatment/control counts by targeter country
  - Total (per targeter): Australia 28, Brazil 28, Canada 36, Chile 26, Colombia 20, Korea 30, Mexico 34, Norway 20, Peru 22, Thailand 22, United Kingdom 66, Total 332
  - Treatment (per targeter): 14, 14, 18, 13, 10, 15, 17, 10, 11, 11, 33, Total 166
  - Controls (per targeter): 14, 14, 18, 13, 10, 15, 17, 10, 11, 11, 33, Total 166
  - Additional country entries (counts as in table): Argentina 1, France 2, Germany 8, Hong Kong SAR 1, India 4, Indonesia 4, Italy 3, Japan 1, Malaysia 2, Netherlands 1, Singapore 1, Spain 1, Switzerland 1, United States 1, Venezuela 1, (and further counts as shown)
- Panel C: Control Group 3 — totals and treatment/control counts by targeter country
  - Total (per targeter): Australia 22, Brazil 21, Canada 31, Chile 17, Colombia 16, Korea 23, Mexico 26, Norway 15, Peru 17, Thailand 19, United Kingdom 45, Total 252
  - Treatment (per targeter): 14, 14, 18, 13, 10, 15, 17, 10, 11, 11, 33, Total 166
  - Controls (per targeter): 8, 7, 13, 4, 6, 8, 9, 5, 6, 8, 12, Total 86
  - Additional country entries (counts as in table): France 2, Germany 8, India 7, Spain 5, United States 13, Venezuela 4, (and other per-cell counts as shown)

### Table 3 — Regression results: specification structure and reported statistics
- Equation estimated: equation (13) (results presented in Table 3).
- Panels A–C present results for the three matching methods (1–3 respectively):
  - Panel A: Nearest Neighbor Matching (with replacement)
  - Panel B: One-for-One Nearest Neighbor Matching (without replacement)
  - Panel C: "Best country" Nearest Neighbor Matching (with replacement)
- Models:
  - Model I: "suppresses the interaction effect, measuring only the levels (or unconditional) effect of IT adoption on forecast errors." (denoted "Levels")
  - Model II: "estimates the full equation using OLS." (denoted "Interaction (OLS)")
  - Model III: "uses the IV strategy (two-stage least squares, 2SLS)." (denoted "Interaction (2SLS)")
- Residual conventions by column: iid, robust, clustered (three columns for each model specification)
- Control: dummy for advanced countries included in each specification; "advanced countries in the dataset are Australia, Canada, France, Germany, Italy, Japan, the Netherlands, Norway, Spain, Switzerland, the United Kingdom, and the United States."

### Selected coefficient estimates and statistics (as reported in Table 3)
- Panel A: Nearest Neighbor Matching (with replacement)
  - IT (Levels / Interaction columns):  1:16 yyy (:272),  1:16 yyy (:305),  1:16  (:666)
  - V_ij0: 0 :153 (:0480),  :153 (:0874),  :153 (:124)
  - IT × V_ij0:  :752 yyy (:119),  :752 yyy (:0891),  :752 yyy (:154)
  - Adv: :0742 (:272), :0742 (:316), :0742 (:690)
  - Const: :193 (:250), :193 (:424), :193 (:864)
  - Other reported figures in Panel A: :200 (:365), :200 (:245), :200 (:423), :831 (:460), :831 yyy (:234), :831 (:421)
  - F_stat entries: 9:49 yyy, 23:0 yyy, 2:47 
  - R2 entries: :0545, :0545, :0545 (and additional R2 values :216, :216, :216; :193, :193, :193)
- Panel B: One-for-One Nearest Neighbor Matching (without replacement)
  - IT:  :961 yyy (:300),  :961  (:336),  :961 (:729)
  - V_ij0:  :286 yyy (:0491),  :286 yyy (:0882),  :286  (:112)
  - IT × V_ij0:  :637 yyy (:128),  :637 yyy (:0891),  :637 yyy (:152)
  - Adv: :309 (:300), :309 (:347), :309 (:753)
  - Const:  :112 (:275),  :112 (:473),  :112 (:979)
  - Additional numeric entries: :0653 (:397), :0653 (:276), :0653 (:489); :589 (:499), :589  (:291), :589 (:462)
  - F_stat entries: 6:24 , 23:0 yyy, 2:37
  - R2 entries: :0366, :0366, :0366; :225, :225, :225; :210, :210, :210
- Panel C: "Best country" Nearest Neighbor Matching (with replacement)
  - IT:  :827  (:335),  :827  (:377),  :827 (1:26)
  - V_ij0:  :365 yyy (:0477),  :365 yyy (:0631),  :365  (:107)
  - IT × V_ij0:  :598 yyy (:139),  :598 yyy (:0705),  :598  (:190)
  - Adv: :453 (:336), :453 (:388), :453 (1:30)
  - Const:  :311 (:306),  :311 (:535),  :311 (1:80)
  - Additional numeric entries: :389 (:440), :389 (:313), :389 (:947); :472 (:550), :472  (:267), :472 (:735)
  - F_stat entries: 4:47 , 23:6 yyy, 2:22
  - R2 entries: :0265, :0265, :0265; :246, :246, :246; :223, :223, :223

### Model estimation notes and notation
- Models reported with three residual treatments per specification: iid, robust, clustered.
- Significance notation: "SigniÖcance levels in this and subsequent tables are denoted as: yyy" (table truncates after that label in the provided content).

*Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06289.pdf*

### 0.1 percent;

### _wp06289 - 0.1 percent;

### Main empirical findings
- Table 3: strong evidence for the conditional effect predicted by the model (columns II and III), but little evidence for an unconditional effect (column I).
  - OLS estimate for b1T (column II): between 0.6 (panel C) and 0.75 (panel A).
  - Using clustered standard errors, OLS estimates are highly statistically significant: at least at the 1% level (panel C) and generally at the 0.1% level (panels A and B).
  - IV/2SLS estimates (column III): point estimate for b1T between 0.8 and 0.94; statistical significance level same as OLS results.
  - The IV strategy eliminates mean reversion present in OLS results: IV estimate of b1 is not significantly different from zero whereas OLS estimate is highly significant and has a much higher point estimate.
- Column I (unconditional effect): not generally statistically significant and carries a negative sign in all specifications.
- Placebo tests (Table 4): shifting IT adoption window 12 months earlier or 12 months later yields no statistically significant conditional or unconditional effect from the placebo IT adoption variable.
  - Note: the first adoption episode (Canada) drops out of the sample due to data constraints.
- Robustness checks (Table 5):
  - Trimming controls with the highest 5 percent of Vij0, or dropping controls with Vij0 above treatment-group maximum (7.4), has little substantial effect on the measured conditional effect of IT.
  - IV estimates on restricted samples often strengthen point estimates and significance for b1T.
  - Dropping Venezuela (majority of outliers) reduces point estimate on b1T but coefficient remains significantly different from zero except in panel C (columns VIII and IX).
  - One case reports an IV p-value of .122 indicating borderline statistical significance.

### Standard errors, clustering, and inference
- Three sets of standard errors presented for each model (in parentheses):
  - Standard OLS/2SLS residuals based on iid homoskedastic variance.
  - Robust standard errors (drops homoskedasticity assumption but maintains independence).
  - Preferred measure: clustered standard errors clustering according to IT adoption episode (1-11) and within each episode by country to control for non-independence across forecasters within each country.
- Rationale for clustering:
  - Bertrand, Du≈o, and Mullainathan (2004) show difference-in-differences with repeated observations within groups overstates significance if within-group correlation is ignored.
  - Clustering (e.g., Stata ìclusterî option) is a simple and effective remedy.
- Number of clusters for clustered residuals reported for panels:
  - Panel A: 58 clusters.
  - Panel B: 63 clusters.
  - Panel C: 23 clusters.
- IV estimates (with clustered residuals) pass:
  - Anderson canonical correlations LR statistic (instrument relevance).
  - Hansenís J statistic (Hansen-Sargan overidentification test for exogeneity).
  - Stock and Yogo tests for weak instruments based on 2SLS bias and size (noting these tests strictly require homoskedastic residuals).

### Quantitative sample and observed data reporting
- Observations (weighted) reported in Table 4: 272 272 272 352 352 352.
- Clusters and Obs reported for Table 5 panels:
  - Panel A: Obs. (Clusters) 307 (54), 304 (54), 305 (53) across specifications.
  - Panel B: Obs. (Clusters) 316 (62), 308 (61), 302 (57).
  - Panel C: Obs. (Clusters) 306 (21), 298 (21), 294 (20).
- In propensity-score estimation:
  - Total number of observations used to estimate ps: 2,141.
  - Probit regression: χ2(8) statistic of 87.6 [p-value=0.000] and a pseudo-R2 of .07.

### Interpretation and conclusions
- The paper tests the proposition derived from a signal-extraction model: if IT enhances transparency as assumed, its introduction should promote convergence to lower forecast errors—i.e., forecast errors should decline under IT proportionately to forecasters' initial errors.
- Empirical conclusion: adoption of IT leads to greater convergence (stronger mean reversion) in inflation forecast errors, as predicted by the model. This conditional effect is robust to:
  - Different propensity score matching strategies (three algorithms).
  - Trimming of control observations with large prior forecast errors.
  - Placebo timing tests.
  - Instrumenting for initial forecast accuracy (IV/2SLS), which strengthens the estimated conditional effect rather than eliminating it.
- There is little or no evidence of an unconditional effect of IT on forecast error levels, confirming Johnson (2002).
- Interpretation aligned with Morris and Shin (2002): the significant conditional effect (b1T < 0) supports the idea that better public information benefits forecasters with poor private information most. No evidence found that IT leads to higher forecast errors for the best forecasters.
- Policy implication: results constitute fairly compelling evidence that IT enhances transparency in inflation forecasting—arguably the central bankís primary objective—thus of interest to policymakers focused on expectations and the monetary transmission mechanism.
- Caveat and suggestion: further research could test IT's transparency effect for other variables; preliminary results suggest little or no effect for output growth forecasting.

### Model comparative statics (Morris and Shin 2002 model summary)
- Setup:
  - Forecasters have utility u_i(f; θ) = (1−r) (f_i − θ)^2 − r L_i − L, where r ∈ [0,1] is weight on ìbeauty contestî element and L_i = ∫_0^1 (f_j − f_i)^2 dj.
  - Public signal θ_C precision denoted β; private signal θ_i precision denoted φ.
- Equilibrium forecast weight:
  - f_i = (β θ_C + φ (1−r) θ_i) / (β + φ (1−r)).
- Expected mean square forecast error:
  - E[(f_i − θ)^2] = (β + φ (1−r)^2) / (β + φ (1−r))^2.
- Comparative statics:
  - ∂E[(f_i − θ)^2]/∂β = − [β − (2r−1) (1−r) φ] / (β + φ (1−r))^3.
  - Predictions:
    - Prediction (i): For the worst forecasters (φ < 8β), better public information always reduces forecast errors.
    - Prediction (ii): If beauty contest element is important (r ≥ 0.5), enhanced transparency can increase forecast errors for the best forecasters when φ > β (2r−1)(1−r).
  - Cross-partial:
    - ∂^2 E[(f_i − θ)^2] / ∂β∂φ = 2 (1−r) [β (1 + r) − (2r−1) (1−r) φ] / (β + φ (1−r))^4.
    - This is ?0 as β/φ ? (2r−1) (1−r) / (1 + r).
  - Svensson (2006) noted realistic parameter values imply β/φ > (2r−1) (1−r), making prediction (ii) unlikely in the data.

### Matching algorithms (propensity-score matching)
- First-stage: estimate propensity score ps via a probit with eight right-hand-side variables:
  - Vij0; g; Vij0; ΔVij−1; ΔVij−1,g; fij0; g; fij0; Δfij−1; g; Δfij−1 (exact variable names preserved as in source).
- Total observations for ps estimation: 2,141; common support required (treatment observations with ps outside control support dropped).
- Algorithm 1: nearest-neighbor matching (with replacement)
  - Random ordering; for each treatment observation, select nearest neighbor in ps within same episode; matching with replacement; weighted to give effective dataset size of 332.
- Algorithm 2: one-to-one matching (without replacement)
  - As Algorithm 1 except matching without replacement producing unique correspondence between 166 control and 166 treatment observations.
- Algorithm 3: nearest-neighbor matching (with replacement) from the ìbestî available country or countries only
  - Step 1 replicates Algorithm 1.
  - Step 2: pick the ìbestî country as that with highest number of frequency-weighted matches from Step 1 (tie-breaker picks both); repeat matching using only forecasters from these best countries as the pool of controls.

*Italic source: _wp06289 - 0.1 percent;*

### REFERENCES

### _wp06289 - REFERENCES

### Inflation targeting and monetary policy frameworks
- Ball, Laurence, and Niamh Sheridan, 2004, “Does Inflation Targeting Matter?” in The Inflation Targeting Debate, ed. by Bernanke and Woodford (Chicago: University of Chicago Press).
- Berg, Claes, 2005, “Experience of Inflation-Targeting in 20 Countries,” Sveriges Riksbank Economic Review, Vol. 1, pp. 20–47.
- Bernanke, Ben, Thomas Laubach, Frederic Mishkin, and Adam Posen, 1999, Inflation Targeting: Lessons from the International Experience (Princeton: Princeton University Press).
- Corbo, Vittorio, Oscar Landerretche, Klaus Schmidt-Hebbel (2001), “Assessing Inflation Targeting after a Decade of World Experience,” International Journal of Finance and Economics 6 (4), pp. 343-368.
- Faust, Jon, and Dale Henderson, 2004, “Is Inflation Targeting Best-Practice Monetary Policy?” Federal Reserve Bank of St. Louis Review, Vol. 86, pp. 117–44.
- Johnson, David, 2002, “The Effect of Inflation Targeting on the Behavior of Expected Inflation: Evidence from an 11 Country Panel,” Journal of Monetary Economics, Vol. 49, pp. 1521–38.
- Kuttner, Kenneth, and Adam Posen, 1999, “Does Talk Matter After All? Inflation Targeting and Central Bank Behavior,” Staff Report 88, Federal Reserve Bank of New York.
- Mahadeva, Lavan, and Gabriel Sterne, 2000, Monetary Policy Frameworks in a Global Context (London: Routledge).
- Mishkin, Frederic, and Klaus Schmidt-Hebbel, 2001, “One Decade of Inflation Targeting in the World: What Do We Know and What Do We Need to Know?” NBER Working Paper No. 8397 (Cambridge, Massachusetts: National Bureau of Economic Research).
- Petursson, Thorarinn, 2004, “The Effects of Inflation Targeting on Macroeconomic Performance,” Central Bank of Iceland Working Paper No. 23.
- Roger, Scott, and Mark Stone, 2005, “On Target? The International Experience with Achieving Inflation Targets,” IMF Working Paper 05/163 (Washington: International Monetary Fund).
- Svensson, Lars, 1999, “Inflation Targeting as a Monetary Policy Rule,” Journal of Monetary Economics, Vol. 43, pp. 607–54.
- Svensson, Lars, “Inflation Targeting: Should it be Modeled as an Instrument Rule or a Targeting Rule?” NBER Working Paper No. 8925 (Cambridge, Massachusetts: National Bureau of Economic Research).
- Vega, Marco, and Diego Winkelried, 2005, “Inflation Targeting and Inflation Behavior: A Successful Story?” International Journal of Central Banking, Vol. 1, No. 3, pp. 153–75.

### Central bank transparency, communication, and credibility
- Blinder, Alan, Charles Goodhart, Philipp Hildebrand, David Lipton, and Charles Wyplosz, 2001, How Do Central Banks Talk? Geneva Report on the World Economy 3 (London: Centre for Economic Policy Research).
- Carpenter, Seth, 2004, “Transparency and Monetary Policy: What Does the Academic Literature Tell Policymakers” (unpublished; Board of Governors of the Federal Reserve System).
- Chortareas, Georgios, David Stasavage, and Gabriel Sterne, 2002, “Does It Pay to be Transparent? International Evidence from Central Bank Forecasts,” Federal Reserve Bank of St. Louis Review, July/August 2002, pp. 99–118.
- Eijffinger, Sylveser, and Petra Geraats, 2006, “How Transparent are Central Banks?” European Journal of Political Economy, Vol. 22, No. 1, pp. 1–21.
- Geraats, Petra, 2002, “Central Bank Transparency,” Economica Journal, Vol. 112, pp. F532–65.
- Goodhart, Charles, and Ellen Meade, 2004, “Central Banks and Supreme Courts,” Moneda y Credito, Vol. 218 (Madrid: Fundacion Santander Central Hispano).
- Meade, Ellen, 2006, “Dissents and Disagreement on the Fed’s FOMC:  Understanding Regional Affiliations and Limits to Transparency,” Netherlands Central Bank Research Department Working Paper 094.
- Romer, Christine, and David Romer, 2000, “Federal Reserve Information and the Behavior of Interest Rates,” American Economic Review, Vol. 90, No. 3, pp. 429–57.
- Kydland, Finn, and Edward Prescott, 1977, “Rules Rather than Discretion: The Inconsistency of Optimal Plans,” Journal of Political Economy, Vol. 85, No. 3, pp. 473–91.
- Cukierman, Alex, and Allan Meltzer, 1986, “A Theory of Ambiguity, Credibility, and Inflation Under Discretion and Asymmetric Information,” Econometrica, Vol. 54, No. 5, pp. 1099–1128.
- Morris, Stephen, and Hyun Song Shin, 2002, “Social Value of Public Information,” American Economic Review, Vol. 92, No. 5, pp. 1521–34.
- Morris, Stephen, Hyun Song Shin, and Hui Tong, 2006, “Social Value of Public Information: Morris and Shin (2002) Is Actually Pro-Transparency, Not Con: Reply,” American Economic Review, Vol. 96, No. 1, pp. 453–55.
- Svensson, Lars, “Social Value of Public Information: Comment: Morris and Shin (2002) “Is Actually Pro-Transparency, not Con,” American Economic Review, Vol. 96, No. 1, pp. 448–52.
- Bernanke et al., 1999, (see above) contains lessons on transparency in the context of inflation targeting.

### Empirical methods, forecasting, and evaluation
- Ang, Andrew, Geert Bakaert, and Min Wai, 2005, “Do Macro Variables, Asset Markets or Surveys Forecast Inflation Better?” NBER Working Paper No. 11538 (Cambridge, Massachusetts: National Bureau of Economic Research).
- Berger, Helge, Michael Ehrmann, and Marcel Fratzscher, 2006, “Forecasting ECB Monetary Policy: Accuracy Is (Still) a Matter of Geography,” IMF Working Paper 06/41 (Washington: International Monetary Fund).
- Gurkaynak, Refet, Andrew Levin, and Eric Swanson, 2005, “Does Inflation Targeting Anchor Long-Run Inflation Expectations? Evidence from Long-Term Bond Yields in the U.S., U.K., and Sweden.” Mimeo.
- Kydland and Prescott, 1977, (see above) on rules versus discretion relevant to policy evaluation.
- Levin, Refet et al., (relevant forecasting/expectations work cited above).
- Leuven, Edwin, and Barbara Sianesi, 2003, “PSMATCH2: Stata Module to Perform Full Mahalanobis and Propensity Score Matching, Common Support Graphing, and Covariate Imbalance Testing.” Available via Internet: http://ideas.repec.org/c.boc/bocode/s432001.html. Version 3.0.0.
- Smith, Jeffrey A., and Petra E. Todd, 2005, “Does Matching Address Lalonde’s Critique of Nonexperimental Estimators?” Journal of Econometrics, Vol. 125, No. 2, pp. 305–53.
- Bertrand, Marianne, Esther Duflo, and Sendhil Mullainathan, 2004, “How Much Should We Trust Differences-in-Differences Estimates?” Quarterly Journal of Economics, Vol. 119, pp. 249–75.
- Besley, Timothy, and Anne Case, 2000, “Unnatural Experiments? Estimating the Incidence of Endogenous Policies,” Economic Journal, Vol. 110, pp. F672–94.
- Stock, James H., Jonathan H. Wright, and Motohiro Yogo, 2002, “A Survey of Weak Instruments and Weak Identification in Generalized Method of Moments,” Journal of Business and Economic Statistics, Vol. 20, No. 4, pp. 518–29.
- Stock, James H., and Motohiro Yogo, 2002, “Testing for Weak Instruments in Linear IV Regression,” NBER Technical Working Paper No. 284 (Cambridge, Massachusetts: National Bureau of Economic Research).

### Foundational and related theoretical works
- Keynes, John Maynard, 1936, The General Theory of Employment, Interest and Money (London: Macmillan).
- Barro, Robert, and David Gordon, 1983, “A Positive Theory of Monetary Policy in a Natural Rate Model,” Journal of Political Economy, Vol. 91, No. 4, pp. 589–610.
- Kydland, Finn, and Edward Prescott, 1977, (see above).
- Cukierman and Meltzer, 1986, (see above).
- Morriss and Shin, 2002 and replies, (see above).
- King, Mervyn, 1997, The Inflation Target Five Years On (lecture delivered at the LSE, October 29 (unpublished; London, Bank of England).

*Source: _wp06289 - REFERENCES*

---


_Source: https://www.imf.org/-/media/websites/imf/imported-full-text-pdf/external/pubs/ft/wp/2006/_wp06289.pdf_